MODIFICATION OF THE FMEA METHOD USING MACHINE LEARNING ALGORITHMS

Abstract

Risk assessment is an important task in any field, from manufacturing to medicine. Risks accompany
a project, product or process throughout its life, from the moment of planning until its
complete termination. Each of them has its own approaches. These include FMEA (Failure Mode
and Effects Analysis) - analysis of the types and consequences of failures. The proposed model is
based on the FMEA method, which is based on risk assessment according to three criteria: the
severity of the consequences when a threat is realized and the complexity of identifying a failure,
the probability of occurrence. The first two criteria are based on expert assessment obtained in
accordance with artificial intelligence methods. The authors proposed a modification of the third
criterion. In our work, we replaced the expert assessment of the “probability of occurrence” criterion
with a machine learning model capable of predicting this indicator based on statistical data.
We carried out the first stage of research into the task at hand on NASA’s open dataset about engine
operating cycles before failure. Initially, the task was set to predict the remaining number of
cycles before failure, then we moved to the classification task, determining whether the equipment
is at risk, depending on its potential remaining life. The best result was obtained by the support
vector machine (SVM), with a classification accuracy of 80%. The goal of the work is to create a
risk assessment model based on the FMEA methodology, which allows to improve the quality of
assessment, reduce subjectivity in decision making, making a forecast based on historical data,
and not just the subjective experience of an expert.